Parallelization Of Adaboost Algorithm And Its Application In Target Classification

Target classification is a key link in Intelligent Video Monitoring and Analysis System, because it should be detected, classified, analyzed and understood during the process of monitoring moving targets. Adaboost(Adaptive Boosting) is a classification algorithm which is the most widely used. Its core thought is using the same training set to train for getting different weak classifiers and then putting these weak classifiers together to form a stronger classifier. In order to get a better performance Adaboost classifier, it often takes a lot of time on training the samples. And this training algorithm requires a large memory space to execute code, so Adaboost algorithm is difficult to run on a normal personal computer.In order to reduce the training time of traditional Adaboost algorithm, this paper used the MIC and GPGPU to optimize the Adaboost:A. Hot spot analysis was carried out on the traditional Adaboost algorithm.We found more than 90% of the time consuming focus on eigenvalue calculation and sorting during training the weak classifier of Adaboost algorithm;B. According to different coprocessors’ hardware architecture and programming style, we used the Bitonic Sort method to optimize sorting, changed the original way of data storage, reduce random visit to save time at the GPGPU platform, which improve the sample training speed. In addition, this article carried out the corresponding parallel optimization experiment of Adaboost using the MIC. We used MPI/Open MP parallel programming tools to optimize the function of heat relatively concentrated. The two optimization strategy —— coarse-grained parallel and the fine-grained parallel were used in it. We test the implementation on a sample set which contains 25,600 samples of size 18*18. The results show that 3.8 time speedup has been achieved at MIC and 7.2 time speedup has been achieved at GPGPU. Experimental results show that GPGPU is superior in terms of processing image data.In order to improve the recognition rate of algorithm, this paper presents a new way of sample set collection for vehicle recognition, which greatly improves the accuracy of the original one. In addition, we also put forward a parallel optimization scheme for speeding up the process of target classification.